OpenLMM Spatial - LEGO-Puzzles (Circular): leaderboard
LEGO-Puzzles from the OpenCompass OpenLMM Spatial Leaderboard (VLMEvalKit): multi-step spatial reasoning over LEGO assembly images, scored with CircularEval (a question counts only if answered correctly under every rotation of its options). Accuracy in percent, higher is better.
Metric: CircularEval Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 58 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 55.5 | #105 |
| 2 | Gemini 2.5 Pro | 49.3 | #159 |
| 3 | GPT-4.1 (2025-04-14) | 43.8 | #257 |
| 4 | GPT-5 Mini (2025-08-07) | 40.3 | #175 |
| 5 | InternVL3-38B | 38.5 | #413 |
| 6 | InternVL3-78B | 37.5 | #361 |
| 7 | Gemini 2.5 Flash | 36.8 | #255 |
| 8 | Claude 3.5 Sonnet (20241022) | 36 | #302 |
| 9 | Claude 3.7 Sonnet | 35.4 | #259 |
| 10 | Qwen 2 VL 72B | 35.4 | #450 |
| 11 | Gemini 2.0 Flash | 35.3 | #351 |
| 12 | InternVL2.5-78B | 32.7 | #434 |
| 13 | Seed 1.6 Flash | 31.5 | #438 |
| 14 | Gemini 1.5 Pro (002) | 31.1 | #368 |
| 15 | GPT-4.1 Mini | 30.5 | #363 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openlmm-spatial-lego-puzzles-circular · How It Works · Data refreshed daily, snapshot 2026-11-05.